展示HN：Llmcc – 用于代码理解的多深度架构图
嗨，HN社区，祝大家新年快乐！<p>我一直在开发llmcc，这是一款能够为代码库生成多种“缩放级别”架构视图的工具。这个想法是：编码代理（和人类）需要从不同的距离理解系统——有时你需要高层次的 crate/模块边界，有时则需要低层次的符号关系。<p>它使用PageRank算法来突出最“重要”的节点，因此你可以在几秒钟内获取一个60万行代码的代码库的聚焦视图。<p>目前仅支持Rust，但架构设计上支持多种语言。<p>之所以开发这个工具，是因为我发现编码代理在大型代码库中常常迷失——要么过于放大（缺乏上下文），要么试图处理过多信息（达到令牌限制）。拥有一个树状视图，能够上下浏览，似乎是一个合适的基本功能。<p>代码库：<a href="https://github.com/allenanswerzq/llmcc" rel="nofollow">https://github.com/allenanswerzq/llmcc</a>，可以查看SVG或基准测试。<p>期待反馈：
- 你们觉得这个工具有用吗？是否愿意继续投入时间？
- 你们希望下一个支持哪些语言？
- 对于这个项目，你们有什么关于编码代理的想法？<p>欢迎提问！
查看原文
Hi, HN community, happy new year to all!<p>I've been working on llmcc, a tool that generates architecture view at multiple "zoom levels" for codebases. The idea: coding agents (and humans) need to understand systems from different distances—sometimes you need the high-level crate/module boundaries, sometimes the low-level symbol relationships.<p>It uses PageRank to surface the most "important" nodes, so you can get a focused view of a 600K LoC codebase in seconds.<p>Currently Rust-only, but the architecture is designed to support multiple languages.<p>Built this because I kept seeing coding agents get lost in large codebases—they'd either get too zoomed in (missing context) or try to process too much (hitting token limits). Having a tree-like view where you can walk up/down seemed like the right primitive.<p>Repo: <a href="https://github.com/allenanswerzq/llmcc" rel="nofollow">https://github.com/allenanswerzq/llmcc</a>, checked out for svgs or benchmark.<p>Looking for feedbacks:
- Do you guys feel could be useful and continue spend time to for it?
- What languages would you want next?
- Any kinds of ideas regarding this project for coding agents?<p>Happy to answer questions!